Sub-scene Regulated Attentional Pooling for Point Cloud Segmentation

Semantic segmentation is one of the key tasks in point cloud processing. To better facilitate the improvement in model design, we conduct a thorough review of various network architecture for general point cloud processing as well as specifically for point cloud semantic segmentation. According to the literature, point-based networks demonstrate great potentials in achieving superior performance due to their ability to directly handle 3D points, compared with multi-view and voxel-based approaches. However, point-based networks usually rely on the max pooling operation to extract useful point features at each of the sub-sampling stages. Though effective, the above-mentioned pooling scheme fails to address the importance of surrounding points, thus increasing the risk of missing useful cues during sub-sampling. In this paper, we propose a novel sub-scene regulated attentional pooling (SRAP) for point cloud segmentation. Specifically, the proposed SRAP devises: 1) an attentional pooling incorporated with the max-pooling to effectively model the relations between sampled points and their neighbour points; and 2) two complementary sub-scene regulations, i.e., intra-consistent regulation to retain the overall sub-scene semantics in point features and inter-diverse regulation to enable discriminativeness between the point features of different sub-scenes. Therefore, the proposed SRAP can effectively improve the point cloud segmentation performance by maintaining the informative point features within sub-scenes. We conduct extensive experiments on two popular point cloud segmentation data sets, Semantic3D and S3DIS. The experimental results demonstrate the significant generalizability of the proposed SRAP in conjunction with different point cloud segmentation methods. For example, SRAP significantly improves the state-of-the-art method, RandLA-Net, from 77.4 to 79.4 on Semantic3D